Research on the Influence of Camera Velocity on Image Blur and a Method to Improve Object Detection Precision

The relative motion between camera and target objects will inevitably result in image blurring, which will cause poor performance of visual perception algorithms. The existing methods focus on coarse-grained classification of random camera jitter to study the degradation of object detection performance. However, there are few studies on blur caused by camera with high-speed movement. In this paper, a novel idea that focuses on the relationship between the performance of the camera's visual perception algorithm (taking object detection as an example) and the forward speed of the camera is proposed. A four-wheeled experimental platform that can move in a straight line at a uniform speed of up to 10km/h was designed. Some sensors such as cameras and speed meters were mounted to take motion blurred pictures at different driving speeds. Then, object detection method YOLO-v5 was performed on the above image data to get the confidence of detection frames at different motion speeds, thus a precise expression of relationship between the confidence and speed could be obtain using polynomial regression method. Further, a preliminarily practice that to improve the object detection performance of this type of motion blur image was conducted. DeblurGAN-v2, a deblurring algorithm based on the generative adversarial networks, is used to deblur the original blurred pictures. By comparing with the result of object detected before and after deblurring, the conclusion that the object detection performance is improved as a whole after deblurring was drawn. These results may help expand the range of image perception (blurred images due to high speed) and generate some potential applications.